def eval_forward(self, x, u):
"""
Evaluates the layer forward, i.e. from input x and random numbers u to output y.
:param x: numpy array
:param u: numpy array
:return: numpy array
"""
if self.eval_forward_f is None:
# conditional input
tt_x = tt.matrix('x')
# masked random numbers
tt_u = tt.matrix('u')
mu = self.mask * tt_u
# scale net
s_net = nn.FeedforwardNet(self.n_inputs + self.n_outputs,
tt.concatenate([tt_x, mu], axis=1))
for h in self.s_hiddens:
s_net.addLayer(h, self.s_act)
s_net.addLayer(self.n_outputs, 'linear')
util.copy_model_parms(self.s_net, s_net)
s = s_net.output
# translate net
t_net = nn.FeedforwardNet(self.n_inputs + self.n_outputs,
tt.concatenate([tt_x, mu], axis=1))
for h in self.t_hiddens:
t_net.addLayer(h, self.t_act)
t_net.addLayer(self.n_outputs, 'linear')
util.copy_model_parms(self.t_net, t_net)
t = t_net.output
# transform (x,u) -> y
y = mu + (1.0 - self.mask) * (tt_u * tt.exp(s) + t)
# compile theano function
self.eval_forward_f = theano.function(inputs=[tt_x, tt_u],
outputs=y)
return self.eval_forward_f(x.astype(dtype), u.astype(dtype))
def eval_forward(self, u):
"""
Evaluates the layer forward, i.e. from random numbers u to output x.
:param u: numpy array
:return: numpy array
"""
if self.eval_forward_f is None:
# masked random numbers
tt_u = tt.matrix('u')
mu = self.mask * tt_u
# scale net
s_net = nn.FeedforwardNet(self.n_inputs, mu)
for h in self.s_hiddens:
s_net.addLayer(h, self.s_act)
s_net.addLayer(self.n_inputs, 'linear')
util.copy_model_parms(self.s_net, s_net)
s = s_net.output
# translate net
t_net = nn.FeedforwardNet(self.n_inputs, mu)
for h in self.t_hiddens:
t_net.addLayer(h, self.t_act)
t_net.addLayer(self.n_inputs, 'linear')
util.copy_model_parms(self.t_net, t_net)
t = t_net.output
# transform u -> x
x = mu + (1.0 - self.mask) * (tt_u * tt.exp(s) + t)
# compile theano function
self.eval_forward_f = theano.function(
inputs=[tt_u],
outputs=x
)
return self.eval_forward_f(u.astype(dtype))